Han and Niranjani work with pi0, a flow-based vision-language-action policy that generates a 350-dimensional action chunk (50 actions of 7 numbers) from images, state and language instruction through 10 denoising steps.
Classifier gradients let frozen robot policies adapt to new bodies without demonstrations
Han and Niranjani steer a frozen flow-based VLA with a lightweight embodiment classifier, raising success from 72% to 80% on modified LIBERO controllers.
Top University
Devin Han · Hershraj Niranjani
UC Berkeley
Research Digest··3 min read
The authors train a small classifier to distinguish action chunks that succeed on a source robot from those that succeed on a modified target body, then use its gradient to guide the frozen pi0 policy during sampling.
§